Interverntion-based Autoencoder

This is an autoencoder that generates a latent representation that is disentangled w.r.t. different interventions performed on the generation of its observations.

Requirements

In order to run the neural network code you will require the following:

python 3.9.7
numpy 1.20.3
torch 1.8.1

In order to generate your own data from our example, you will also require the following:

scipy 1.7.1

You may run the code with different versions, but these are the versions we have verified.

Architecture

TBA

Get Started

In order to train your own intervention-based autoencoder (iAE) you will need to generate your own data or use the data from our example.

Train iAE with our data

You can start by running the iAE on the data from our example. Our example can be found in /examples/state_tomography.py. Here, we generate data from different types of measurements on a random two-qubit quantum state. Measurements are fixed and either act on qubit 1, qubit 2 or both qubits. Each datapoint corresponds to a newly created random mixed state.

The data is separated into three different types of intervention:

Remove all measurements that affect qubit 2
Remove all measurements that affect qubit 1
No intervention

To generate this data set for different interventions go to /examples and run in your terminal:

python state_tomography.py

After it has finished, you will see 6 files in the folder /data. The following three files contain the training data for the different interventions:

tomography_0.pth
tomography_1.pth
tomography_2.pth

Now you are ready to train the iAE on this data! Therefor, go back to the main folder. Since we have already hyperparamaterized the code for you, you don't need to bother about the paramters in Config.py for now. Just run the following in your terminal to train your iAE:

python main.py

Now, you should start seeing text that is printed to your console. When it has finished, you can review the results in /results. There, you will find four files:

results_loss_rec.txt: This is the recreation loss over time for each intervention.
results_loss.txt: This is the overall loss over time.
results_min.txt: This is the minimization filter over time.
results_sel.txt: This is the selection filters of each intervention over time.

The filters will be of particular interest. You can see here which neurons of the latent representation have been filtered. A value > 0 indicates that a neuron is ignored by the decoder(s).

If everything went well, you should see in results_min.txt something that looks like this:

[-3.0184876918792725, 1.0279388427734375, -3.338914632797241, -4.57394552230835, 1.4305622577667236, -4.4977922439575195, -3.226980209350586, -3.2366509437561035, -4.525698184967041, 1.3006442785263062, -3.021491527557373, -2.9364583492279053, -3.0780858993530273, -4.466502666473389, 0.010970002971589565, -4.533066272735596, -3.0983939170837402, -3.294105291366577, 1.2918269634246826, -4.566442489624023]

This is the global "filter value" for each of the 20 starting neurons. From the last entry, you can see that 15 neurons are still active (indicated by a value < 0).

Similarly, you can analyze the results in results_sel.txt where entries like this should populate the file:

[Intervention: 0], [5.63159704208374, 4.369762897491455, 5.476593494415283, 3.6281468868255615, 3.538800001144409, -3.7709128856658936, 4.755659103393555, 4.2118096351623535, 5.206002235412598, 4.347437381744385, 5.386545181274414, 5.070257663726807, 3.712785005569458, 4.8540778160095215, 4.232072830200195, -3.8091671466827393, 5.3681535720825195, 5.737484931945801, 4.310754299163818, -3.8430206775665283]
[Intervention: 1], [3.563845634460449, 3.504835367202759, 4.641702651977539, -3.869236946105957, 3.972226858139038, 5.328145980834961, 4.377258777618408, 3.7198026180267334, -3.8097751140594482, 3.416163444519043, 4.593560695648193, 3.4797816276550293, 3.712146759033203, -3.751715660095215, 3.3567497730255127, 4.810200214385986, 3.1704533100128174, 4.839340686798096, 3.2992093563079834, 4.509983539581299]
[Intervention: 2], [-2.6669344902038574, 3.1874544620513916, -2.984987497329712, -3.9072558879852295, 2.919259548187256, -3.827608823776245, -2.886353015899658, -2.8864145278930664, -3.8482534885406494, 3.3910930156707764, -2.6667697429656982, -2.590120792388916, -2.728029727935791, -3.803149461746216, 2.3931884765625, -3.864940643310547, -2.757671356201172, -2.9479329586029053, 3.1411244869232178, -3.896395206451416]

These are the "filter values" of each decoder for each of the 20 latent neurons. Here, we can see that intervention 0 and 1 have only 3 disjoint active neurons (corresponding to the 3 variables required for a one-qubit Bloch representation). At the same time, the last decoder, that does not do any intervention, requires all information (15 neurons) to reproduce the mixed states.

Create your own data